Published on

  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide


  1. 1. IEEE TRANSACnONS ON SYSTEMS, M N AND CYBERNETICS, V O L A. 21, NO. 3. MAYNUNE 1991 473 Outline for a Theory o f Intelligence James S. Albus Absfmct -Intelligence i s defined as that which produces SUC- adds information about mental development, emotions, andcessful behavior. Intelligence i s assumed to result from natural behavior.selection. A model is proposed that integrates knowledge from Research in learning automata, neural nets, and brain mod-research in both natural and artificial systems. The model con- eling has given insight into learning and the similaritiessists o f a hierarchical system architecture wherein: 1 control )bandwidth decreases about an order of magnitude a t each higher and differences between neuronal and electronic comput -level, 2) perceptual resolution o f spatial and temporal patterns ing processes. Computer science and artificial intelligencecontracts about an order-of-magnitude at each higher level, 3) is probing the nature of language and image understanding,goals expand in scope and planning horizons expand in space and has made significant progress in rule based reasoning,and time about an order-of-magnitude at each higher level, and4) models of the world and memories of events expand their planning, and problem solving. Game theory and operationsrange in space and time by about an order-of-magnitude at research have developed methods for decision making ineach higher level. At each level, functional modules perform the face of uncertainty. Robotics and autonomous vehiclebehavior generation (task decomposition planning and execution), research has produced advances in real-time sensory process -world modeling, sensory processing, and value judgment. Sensory ing, world modeling, navigation, trajectory generation, andfeedback contml o s are closed at every level. lp o obstacle avoidance. Research in automated manufacturing and process control has produced intelligent hierarchical controls, I INTR ODUCTI ON . distributed databases, representations of object geometry andM UCH IS UNKNOWN about intelligence, and much w remain beyond human comprehension for a very i llong time. The fundamental nature o f intelligence is only material properties, data driven task sequencing, network com- munications, and multiprocessor operating systems. Modern control theory has developed precise understanding of stability,dimly understood, and the elements of self consciousness, adaptability, and controllability under various conditions ofperception, reason, emotion, and intuition are cloaked in feedback and noise. Research in sonar, radar, and optical signalmystery that shrouds the human psyche and fades into the processing has developed methods for fusing sensory inputreligious. Even the definition of intelligence remains a subject from multiple sources, and assessing the believability of noisyof controversy, and so must any theory that attempts to data.explain what intelligence is, how i t originated, or what are Progress i s rapid, and there exists an enormous and rapidlythe fundamental processes by which it functions. growing literature in each of the previous fields. What i s Yet, much is known, both about the mechanisms and func- lacking i s a general theoretical model of intelligence that tiestion of intelligence. The study of intelligent machines and the all these separate areas o f knowledge into a unified framework.neurosciences are both extremely active fields. Many millions This paper is an attempt to formulate at least the broad outlinesof dollars per year are now being spent in Europe, Japan, of such a model.and the United States on computer integrated manufacturing, The ultimate goal is a general theory of intelligence thatrobotics, and intelligent machines for a wide variety of military encompasses both biological and machine instantiations. Theand commercial applications. Around the world, researchers in model presented here incorporates knowledge gained fromthe neurosciences are searching for the anatomical, physiolog - many different sources and the discussion frequently shiftsical, and chemical basis of behavior. back and forth between natural and artificial systems. For Neuroanatomy has produced extensive maps of the inter- example, the definition of intelligence in Section I1 addressesconnecting pathways making up the structure of the brain. both natural and artificial systeqs. Section I11 treats the originNeurophysiology is demonstrating how neurons compute func- and function of intelligence from the standpoint of biological tions and communicate information. Neuropharmacology is evolution. In Section IV, both natural and artificial systemdiscovering many of the transmitter substances that modify elements are discussed. The system architecture describedvalue judgments, compute reward and punishment, activate in Sections V-VI1 derives almost entirely from research inbehavior, and produce learning. Psychophysics provides many robotics and control theory for devices ranging from underseaclues as to how individuals perceive objects, events, time, vehicles to automatic factories. Sections VIII-XI on behaviorand space, and how they reason about relatianships between generation, Sections XI1 and XI11 on world modeling, andthemselves and the external world. Behavipral psychology Section XIV on sensory processing are elaborations of the Manuscript received March 16, 1990; revised November 16, 1990. system architecture o f Section V- VII. These sections all con- The author is with the Robot Systems Division Center for Manufacturing tain numerous references to neurophysiological, psychological,Engineering, National Institute of Standards and Technology, Gaithersburg,MD 20899. and psychophysical phenomena that support the model, and IEEE Log Number 9042583. frequent analogies are drawn between biological and artificial 0018-9472/91/0500 -0473f01.00 0 1991 IEEE
  2. 2. 474 IEEE TRANSACTIONS ON SYSTEMS, MAN. AND CYBERNETICS, VOL. 21. NO. 3. MAY/JUNE 1991 systems. T h e value judgments, described in Section XV, are emotion, and behavior in a sensing, perceiving, knowing,mostly based on the neurophysiology of the limbic system and caring, planning, acting system that can succeed in achieving the psychology of emotion. Section XVI on neural computa - its goals in the world. tion and Section XVlI on learning derive mostly from neural For the purposes of this paper, intelligence will be definednet research. as the ability of a system to act appropriately in an uncertain The model is described in terms of definitions, axioms, environment, where appropriate action is that which increasestheorems, hypotheses, conjectures, and arguments in support the probability o f success, and success is the achievement ofof them. Axioms are statements that are assumed to be true behavioral subgoals that support the system’s ultimate goal.without proof. Theorems are statements that the author feels Both the criteria of success and the systems ultimate goalcould be demonstrated true by existing logical methods or are defined external to the intelligent system. For an intelligentempirical evidence. Few of the theorems are proven, but each machine system, the goals and success criteria are typicallyis followed by informal discussions that support the theorem defined by designers, programmers, and operators. For intelli -and suggest arguments upon which a formal proof might gent biological creatures, the ultimate goal is gene propagation,be constructed. Hypotheses are statements that the author and success criteria are defined by the processes of naturalfeels probably could be demonstrated through future research. selection.Conjectures are statements that the author feels might be Theorem: There are degrees, or levels, of intelligence,demonstrable. and these are determined by: 1) the computational power o f the system’s brain (or computer), 2) the sophistication of algorithms the system uses for sensory processing, world 11. DEFI ITI N OF N O INTELLIGENCE modeling, behavior generating, value judgment, and global In order to be useful in the quest for a general theory, the communication, and 3) the information and values the systemdefinition of intelligence must not be limited to behavior that has stored in its not understood. A useful definition of intelligence should Intelligence can be observed to grow and evolve, bothspan a wide range of capabilities, from those that are well through growth in computational power, and through accu-understood, to those that are beyond comprehension. I t should mulation of knowledge of how to sense, decide, and act in ainclude both biological and machine embodiments, and these complex and changing world. In artificial systems, growth inshould span an intellectual range from that of an insect to computational power and accumulation of knowledge derivesthat of an Einstein, from that of a thermostat to that of the mostly from human hardware engineers and software program -most sophisticated computer system that could ever be built. mers. In natural systems, intelligence grows, over the lifetimeThe definition of intelligence should, for example, include the of an individual, through maturation and learning; and overability of a robot to spotweld an automobile body, the ability intervals spanning generations, through evolution.of a bee to navigate in a field of wild flowers, a squirrel to Note that learning is not required in order to be intelligent,jump from limb to limb, a duck to land in a high wind, and only to become more intelligent as a result of experience.a swallow to work a field of insects. I t should include what Learning is defined as consolidating short-term memory intoenables a pair of blue jays to battle in the branches for a long-term memory, and exhibiting altered behavior because ofnesting site, a pride of lions to pull down a wildebeest, a flock what was remembered. In Section X, learning is discussed asof geese to migrate south in the winter. I t should include what a mechanism for storing knowledge about the external world,enables a human to bake a cake, play the violin, read a book, and for acquiring skills and knowledge of how to act. I t is,write a poem, fight a war, or invent a computer. however, assumed that many creatures can exhibit intelligent At a minimum, intelligence requires the ability to sense the behavior using instinct, without having learned anything.environment, to make decisions, and to control action. Higherlevels of intelligence may include the ability to recognizeobjects and events, to represent knowledge in a world model, 111. T H E ORIGIN AND FUNC ION INTELLIGENCE T OFand to reason about and plan for the future. In advanced forms, Theorem: Natural intelligence, like the brain in which i tintelligence provides the capacity to perceive and understand, appears, is a result of the process of natural choose wisely, and to act successfully under a large variety The brain is first and foremost a control system. Its primaryof circumstances so as to survive, prosper, and reproduce in a function is to produce successful goal-seeking behavior in find-complex and often hostile environment. ing food, avoiding danger, competing for territory, attracting From the viewpoint o f control theory, intelligence might sexual partners, and caring for offspring. Abrains that ever l lbe defined as a knowledgeable “helmsman of behavior”. existed, even those o f the tiniest insects, generate and controlIntelligence i s the integration o f knowledge and feedback behavior. Some brains produce only simple forms of behavior, into a sensory -interactive goal-directed control system that can while others produce very complex behaviors. Only the most make plans, and generate effective, purposeful action directed recent and highly developed brains show any evidence o f toward achieving them. abstract thought. From the viewpoint of psychology, intelligence might be Theorem: For each individual, intelligence provides a mech- defined as a behavioral strategy that gives each individual a anism for generating biologically advantageous behavior. means for maximizing the likelihood of propagating its own Intelligence improves an individual’s ability to act effec - genes. Intelligence is the integration o f perception, reason, tively and choose wisely between alternative behaviors. A l
  3. 3. ALBUS: OUTLINE FOR A THEORY OF INTELLIGENCE 475 else being equal, a more intelligent individual has many Words may be represented by symbols. Syntax, or grammar, advantages over less intelligent rivals in acquiring choice i s the set of rules for generating strings of symbols that territory, gaining access to food, and attracting more desirable form sentences. Semantics is the encoding of information into breeding partners. The intelligent use of aggression helps meaningful patterns, or messages. Messages are sentences that to improve an individual’s position in the social dominance convey useful information. hierarchy. Intelligent predation improves success in capturing Communication requires that information be: 1) encoded,prey. Intelligent exploration improves success in hunting and 2) transmitted, 3) received, 4) decoded, and 5) understood. establishing territory. Intelligent use o f stealth gives a predator Understanding implies that the information in the message has the advantage of surprise. Intelligent use of deception improves been correctly decoded and incorporated into the world modelthe prey’s chances of escaping from danger. of the receiver. Higher levels of intelligence produce capabilities in the Communication may be either intentional or unintentional.individual for thinking ahead, planning before acting, and Intentional communication occurs as the result of a senderreasoning about the probable results of alternative actions. executing a task whose goal i t is to alter the knowledge or be-These abilities give to the more intelligent individual a com- havior of the receiver to the benefit of the sender. Unintentionalpetitive advantage over the less intelligent in the competition communication occurs when a message is unintentionally sent,for survival and gene propagation. Intellectual capacities and or when an intended message is received and understood bybehavioral skills that produce successful hunting and gathering someone other than the intended receiver. Preventing an enemyof food, acquisition and defense of territory, avoidance and from receiving and understanding communication betweenescape from danger, and bearing and raising offspring tend to friendly agents can often be crucial to passed on to succeeding generations. Intellectual capabili - Communication and language are by no means unique toties that produce less successful behaviors reduce the survival human beings. Virtually all creatures, even insects, commu-probability of the brains that generate them. Competition nicate in some way, and hence have some form of language.between individuals thus drives the evolution of intelligence For example, many insects transmit messages announcing theirwithin a species. identity and position. This may be done acoustically, by smell, Theorem: For groups of individuals, intelligence provides or by some visually detectable display. The goal may be toa mechanism for cooperatively generating biologically advan- attract a mate, or to facilitate recognition mYor location bytageous behavior. other members of a group. Species of lower intelligence, such T h e intellectual capacity to simply congregate into flocks, as insects, have very little information to communicate, andherds, schools, and packs increases the number of sensors hence have languages with only a few of what might be calledwatching for danger. The ability to communicate danger words, with little or no grammar. In many cases, languagesignals improves the survival probability of all individuals vocabularies include motions and gestures (i.e., body or signin the group. Communication is most advantageous to those language) as well as acoustic signals generated by variety ofindividuals who are the quickest and most discriminating mechanisms from stamping the feet, to snorts, squeals, chirps,in the recognition of danger messages, and most effective cries, and responding with appropriate action. The intelligence to Theorem: In any species, language evolves to support thecooperate in mutually beneficial activities such as hunting and complexity of messages that can be generated by the intelli -group defense increases the probability of gene propagation gence of that species.for all members o f the group. Depending on its complexity, a language may be capable of A else being equal, the most intelligent individuals and l communicating many messages, or only a few. More intelli-groups within a species will tend to occupy the best territory, gent individuals have a larger vocabulary, and are quicker tobe the most successful in social competition, and have the understand and act on the meaning of chances for their offspring surviving. A else being equal, l Theorem: To the receiver, the benefit, or value, of commu-more intelligent individuals and groups will win out in serious nication i s roughly proportional to the product of the amount ofcompetition with less intelligent individuals and groups. information contained in the message, multiplied by the ability Intelligence is, therefore, the product of continuous com- of the receiver to understand and act on that information,petitive struggles for survival and gene propagation that has multiplied by the importance o f the act to survival and gene taken place between billions of brains, over millions of years. propagation of the receiver. To the sender, the benefit is theThe results of those struggles have been determined in large value of the receiver’s action to the sender, minus the danger measure by the intelligence of the competitors. incurred by transmitting a message that may be intercepted by, and give advantage to, an enemy. Greater intelligence enhances both the individual’s and theA. Communication and Language group’s ability to analyze the environment, to encode and Definition: Communication is the transmission of informa - transmit information about it, to detect messages, to recognizetion between intelligent systems. their significance, and act effectively on information received. Definition: Language is the means by which information is Greater intelligence produces more complex languages capableencoded for purposes of communication. of expressing more information, i.e., more messages with more Language has three basic components: vocabulary, syntax, shades of meaning. ,and semantics. Vocabularv is the set of words in the language. In social soecies. communication also orovides the basis
  4. 4. 476 IEEE TRANSACTIONS ON SYSTEMS. M N AND CYBERNETICS. VOL. 21, NO. 3. MAYiJUNE 199 A, for societal organization. Communication of threats that warn legs, hands, and eyes. Actuators generate forces to point of aggression can help to establish the social dominance sensors, excite transducers, move manipulators, handle tools, hierarchy, and reduce the incidence of physical harm from steer and propel locomotion. An intelligent system may have fights over food, territory, and sexual partners. Communication tens, hundreds, thousands, even millions of actuators, all of of alarm signals indicate the presence of danger, and in some which must be coordinated in order to perform tasks and cases, identify its type and location. Communication of pleas accomplish goals. Natural actuators are muscles and glands. for help enables group members to solicit assistance from one Machine actuators are motors, pistons, valves, solenoids, andanother. Communication between members o f a hunting pack transducers.enable them to remain in formation while spread far apart, and 2) Sensors: Input to an intelligent system is produced byhence to hunt more effectively by cooperating as a team in the sensors, which may include visual brightness and color sen-tracking and killing o f prey. sors: tactile, force, torque, position detectors; velocity, vibra- Among humans, primitive forms of communication include tion, acoustic, range, smell, taste, pressure, and temperaturefacial expressions, cries, gestures, body language, and pan- measuring devices. Sensors may be used to monitor bothtomime. The human brain is, however, capable o f generating the state of the external world and the internal state of theideas of much greater complexity and subtlety than can be intelligent system itself. Sensors provide input to a sensoryexpressed through cries and gestures. In order to transmit mes- processing system.sages commensurate with the complexity of human thought, 3) Sensory Processing: Perception takes place in a sensoryhuman languages have evolved grammatical and semantic processing system element that compares sensory observationsrules capable of stringing words from vocabularies consisting with expectations generated by an internal world model.of thousands of entries into sentences that express ideas Sensory processing algorithms integrate similarities and dif-and concepts with exquisitely subtle nuances of meaning. To ferences between observations and expectations over timesupport this process, the human vocal apparatus has evolved and space so as to detect events and recognize features,complex mechanisms for making a large variety o f sounds. objects, and relationships i n the world. Sensory input data from a wide variety o f sensors over extended periods ofB. Human Intelligence and Technology time are fused into a consistent unified perception of the state of the world. Sensory processing algorithms compute Superior intelligence alone made man a successful hunter. distance, shape, orientation, surface characteristics, physicalThe intellectual capacity to make and use tools, weapons, and dynamical attributes of objects and regions of space.and spoken language made him the most successful of all Sensory processing may include recognition of speech andpredators. In recent millennia, human levels of intelligence interpretation of language and music.have led to the use of fire, the domestication o f animals, 4) World Model: The world model is the intelligent sys-the development of agriculture, the rise of civilization, the tem’s best estimate of the state of the world. The world modelinvention of writing, the building of cities, the practice of includes a database o f knowledge about the world, plus awar, the emergence of science, and the growth of industry. database management system that stores and retrieves infor -These capabilities have extremely high gene propagation value mation. The world model also contains a simulation capabilityfor the individuals and societies that possess them relative to that generates expectations and predictions. T h e world modelthose who do not. Intelligence has thus made modem civilized thus can provide answers to requests for information abouthumans the dominant species on the planet Earth. the present, past, and probable future states of the world. T h e For an individual human, superior intelligence i s an asset in world model provides this information service to the behaviorcompeting for position in the social dominance hierarchy. I t generation system element, so that i t can make intelligentconveys advantage for attracting and winning a desirable mate, plans and behavioral choices, to the sensory processing systemin raising a large, healthy, and prosperous family, and seeing to element, in order for i t to perform correlation, model matching,i t that one’s offspring are well provided for. I n competition be- and model based recognition of states, objects, and events, andtween human groups, more intelligent customs and traditions, to the value judgment system element in order for i t to computeand more highly developed institutions and technology, lead to values such as cost, benefit, risk, uncertainty, importance,the dominance of culture and growth of military and political attractiveness, etc. T h e world model is kept up-to-date by thepower. Less intelligent customs, traditions, and practices, and sensory processing system element.less developed institutions and technology, lead to economicand political decline and eventually to the demise of tribes, 5) Value Judgment: T h e value judgment system elementnations, and civilizations. determines what is good and bad, rewarding and punishing, important and trivial, certain and improbable. The value judg- ment system evaluates both the observed state of the world Iv. THE ELEMENTS OF INTELLIGENCE and the predicted results of hypothesized plans. I t computes Theorem: There are four system elements of intelligence: costs, risks, and benefits both o f observed situations and ofsensory processing, world modeling, behavior generation, and planned activities. It computes the probability of correctnessvalue judgment. Input to, and output from, intelligent systems and assigns believability and uncertainty parameters to stateare via sensors and actuators. variables. I t also assigns attractiveness, or repulsiveness to I Actuators: Output from an intelligent system is produced ) objects, events, regions of space, and other creatures. Theby actuators that move, exert forces, and position arms, value judgment system thus provides the basis for making
  5. 5. ALBUS. OUTLINE FOR A THEORY OF INTELLIGENCE 471 Situruon Planning and decisions -for choosing one action as opposed to another, - Aswrvncnr 1 Exsutiw1 - or for pursuing one object and fleeing from another. Without value judgments, any biological creature would soon be eaten by others, and any artificially intelligent system would soon be disabled by its own inappropriate actions. 6) Behavior Generation: Behavior results from a behavior generating system element that selects goals, and plans and ex - ecutes tasks. Tasks are recursively decomposed into subtasks, and subtasks are sequenced so as to achieve goals. Goals are selected and plans generated by a looping interaction between COMWNDEDbehavior generation, world modeling, and value judgmentelements. The behavior generating system hypothesizes plans, the world model predicts the results o f those plans, and the value judgment element evaluates those results. The behaviorgenerating system then selects the plans with the highest INTERNALevaluations for execution. The behavior generating systemelement also monitors the execution of plans, and modifiesexisting plans whenever the situation requires. Each of the system elements o f intelligence are reasonablywell understood. The phenomena of intelligence, however,requires more than a set of disconnected elements. Intelligencerequires an interconnecting system architecture that enables Fig. 1. Elements o f intelligence and the functional relationships between them.the various system elements to interact and communicate witheach other in intimate and sophisticated ways. A system architecture i s what partitions the system elements Telerobotic Servicer [14] and the Air Force Next Generationof intelligence into computational modules, and interconnects Controller.the modules in networks and hierarchies. I t is what enables the The proposed system architecture organizes the elements ofbehavior generation elements to direct sensors, and to focus intelligence so as to create the functional relationships andsensory processing algorithms on objects and events worthy information flow shown in Fig. 1. In all intelligent systems,of attention, ignoring things that are not important to current a sensory processing system processes sensory information togoals and task priorities. I t is what enables the world model acquire and maintain an internal model of the external answer queries from behavior generating modules, and In a l l systems, a behavior generating system controls actuatorsmake predictions and receive updates from sensory processing so as to pursue behavioral goals in the context of the perceivedmodules. I t is what communicates the value state -variables that world model. In systems o f higher intelligence, the behaviordescribe the success of behavior and the desirability of states generating system element may interact with the world modelof the world from the value judgment element to the goal and value judgment system to reason about space and time,selection subsystem. geometry and dynamics, and to formulate or select plans based on values such as cost, risk, utility, and goal priorities. T h e sensory processing system element may interact with the worldv. A PROPOSED A R C H ITE C ~~ RE INTELLIGENT SYSTEMS model and value judgment system to assign values to perceived FOR A number of system architectures for intelligent machine entities, events, and have been conceived, and a few implemented. [ 11-[15] The proposed system architecture replicates and distributesThe architecture for intelligent systems that wl be proposed the relationships shown in Fig. 1 over a hierarchical computing ilhere is largely based on the real -time control system (RCS) that structure with the logical and temporal properties illustratedhas been implemented in a number of versions over the past 13 in Fig. 2. On the left is an organizational hierarchy whereinyears at the National Institute for Standards and Technology computational nodes are arranged in layers like command(NIST, formerly NBS). RCS was first implemented by Barbera posts in a military organization. Each node in the organiza -for laboratory robotics in the mid 1970’s [7] and adapted by tional hierarchy contains four types of computing modules:Albus, Barbera, and others for manufacturing control in the behavior generating (BG), world modeling (WM), sensoryNIST Automated Manufacturing Research Facility (AMRF) processing (SP), and value judgment (VJ) modules. Eachduring the early 1980’s [ll], Since 1986, RCS has been chain of command in the organizational hierarchy, from each [12].implemented for a number of additional applications, including actuator and each sensor to the highest level of control, canthe NBSDARPA Multiple Autonomous Undersea Vehicle be represented by a computational hierarchy, such as i s shown(MAUV) project [13], the Army Field Material Handling in the center of Fig. 2.Robot, and the Army TMAP and TEAM semiautonomous land At each level, the nodes, and computing modules withinvehicle projects. RCS also forms the basis o f the NASAPJBS the nodes, are richly interconnected to each other by a com-Standard Reference Model Telerobot Control System Archi - munications system. Within each computational node, thetecture (NASREM) being used on the space station Flight communication system provides intermodule communications
  6. 6. 478 IEEE TRANSACTIONS ON SYSTEMS, MN A. AND CYBERNETICS, VOL. 21. NO. 3. MAYiJUNE 1991 COMPUTATIONAL HIERARCHY ORGANIZATIONAL BEHAVIORAL HERARCHY Srmq Vdue Jrdgmenl Lkhrrbr HIERARCHY y WwMHodr11iq Cnmllng Fig. 2. Relationships in hierarchical control systems, On the left is an organizational hierarchy consisting of a tree of command centers, each of which possesses one supervisor and one or more subordinates. In the center i s a computational hierarchy consisting of BG, WM, SP, and VJ modules. Each actuator and each sensors is serviced by a computational hierarchy. On the right is a behavioral hierarchy consisting of trajectories through state-time-space. Commands at a each level can be represented by vectors, or points i n state-space. Sequences of commands and be represented as trajectories through sfate-time-space.of the type shown in Fig. 1. Queries and task status are tions functions may be a computer bus, a local area network,communicated from BG modules to WM modules. Retrievals a common memory, a message passing system, or someof information are communicated from WM modules back to combination thereof. In either biological or artificial systems,the BG modules making the queries. Predicted sensory data i s the communications system may include the functionalitycommunicated from WM modules to SP modules. Updates to of a communications processor, a file server, a databasethe world model are communicated from SP to W M modules. management system, a question answering system, or anObserved entities, events, and situations are communicated indirect addressing or list processing engine. In the systemfrom SP to VJ modules. Values assigned to the world model architecture proposed here, the input/output relationships of therepresentations of these entities, events, and situations are communications system produce the effect of a virtual globalcommunicated from VJ to W M modules. Hypothesized plans memory, or blackboard system [15].are communicated from BG to W M modules. Results are The input command string to each of the BG modulescommunicated from W M to VJ modules. Evaluations are at each level generates a trajectory through state-space ascommunicated from VJ modules back to the BG modules that a function of time. The set of all command strings createhypothesized the plans. a behavioral hierarchy, as shown on the right o f Fig. 2. T h e communications system also communicates between Actuator output trajectories (not shown in Fig. 2) correspondnodes at different levels. Commands are communicated down- to observable output behavior. A the other trajectories in the lward from supervisor BG modules in one level to subordinate behavioral hierarchy constitute the deep structure of behaviorBG modules in the level below. Status reports are commu -nicated back upward through the world model from lowerlevel subordinate BG modules to the upper level supervisorBG modules from which commands were received. Observed VI. HIERARCHICAL VERSUS HORIZONTALentities, events, and situations detected by SP modules at one Fig. 3 shows the organizational hierarchy in more detail,level are communicated upward to SP modules at a higher and illustrates both the hierarchical and horizontal relation -level. Predicted attributes of entities, events, and situations ships involved in the proposed architecture. The architecturestored in the W M modules at a higher level are communi - is hierarchical in that commands and status feedback flowcated downward to lower level W M modules. Output from hierarchically up and down a behavior generating chain ofthe bottom level BG modules is communicated to actuator command. T h e architecture is also hierarchical in that sensorydrive mechanisms. Input to the bottom level SP modules is processing and world modeling functions have hierarchicalcommunicated from sensors. levels of temporal and spatial aggregation. T h e communications system can be implemented in a va- The architecture is horizontal in that data is shared hori-riety of ways. In a biological brain, communication is mostly zontally between heterogeneous modules at the same level.via neuronal axon pathways, although some messages are At each hierarchical level, the architecture is horizontallycommunicated by hormones carried in the bloodstream. In interconnected by wide-bandwidth communication pathwaysartificial systems, the physical implementation o f communica - between BG, WM, SP, and VJ modules in the same node,
  7. 7. ALBUS: OUTLINE FOR A THEORY OF INTELLIGENCEFig. 3. An organization of processing nodes such that the BG modules forma command tree. On the right are examples or the functional characteristicof the BG modules at each level. On the left are examples of the type ofvisual and acoustical entities recognized by the SP modules at each level. Inthe center of level 3 are the type of subsystems represented by processingnodes at level 3.and between nodes at the same level, especially within the Fig. 4. Each layer of the system architecture contains a number of nodes, each of which contains BG, WM, SP, and VJ modules, The nodes aresame command subtree. The horizontal flo:;. of information interconnected as a layered graph, or lattice, through the communicationis most voluminous within a single node, less so between system. Note that the nodes are richly but not fully, interconnected. Outputsrelated nodes in the same command subtree, and relatively low from the bottom layer BG modules drive actuators. Inputs to the bottom layer SP modules convey data from sensors. During operation, goal drivenbandwidth between computing modules in separate command communication path selection mechanisms configure this lattice structure intosubtrees. Communications bandwidth is indicated in Fig. 3 by the organization tree shown in Fig. 3.the relative thickness of the horizontal connections. The volume o f information flowing horizontally within a During operation, goal driven switching mechanisms in the BGsubtree may be orders of magnitude larger than the amount modules (discussed in Section X) assess priorities, negotiateflowing vertically in the command chain. The volume of in- for resources, and coordinate task activities so as to selectformation flowing vertically in the sensory processing system among the possible communication paths o f Fig. 4. As acan also be very high, especially in the vision system. result, each BG module accepts task commands from only The specific configuration of the command tree is task one supervisor at a time, and hence the BG modules form adependent, and therefore not necessarily stationary in time. command tree at every instant in time.Fig. 3 illustrates only one possible configuration that may T h e SP modules are also organized hierarchically, but asexist at a single point in time. During operation, relationships a layered graph, not a tree. At each higher level, sensorybetween modules within and between layers o f the hierarchy information is processed into increasingly higher levels o fmay be reconfigured in order to accomplish different goals, pri- abstraction, but the sensory processing pathways may branchorities, and task requirements. This means that any particular and merge in many different ways.computational node, with its BG, WM, SP, and VJ modules, may belong to one subsystem at one time and a different subsystem a very short time later. For example, the mouth may VII. HIERARCHICALLEVELS be part of the manipulation subsystem (while eating) and the Levels in the behavior generating hierarchy are defined by communication subsystem (while speaking). Similarly, an arm temporal and spatial decomposition of goals and tasks into may be part of the manipulation subsystem (while grasping) levels of resolution. Temporal resolution is manifested in terms and part of the locomotion subsystem (while swimming or of loop bandwidth, sampling rate, and state-change intervals. climbing). Temporal span is measured by the length of historical traces In the biological brain, command tree reconfiguration can and planning horizons. Spatial resolution is manifested in the be implemented through multiple axon pathways that exist, branching of the command tree and the resolution of maps. but are not always activated, between BG modules at dif- Spatial span is measured by the span of control and the range ferent hierarchical levels. These multiple pathways define a of maps. layered graph, or lattice, o f nodes and directed arcs, such as Levels in the sensory processing hierarchy are defined by shown in Fig. 4. They enable each BG module to receive temporal and spatial integration of sensory data into levels of input messages and parameters from several different sources. aggregation. Spatial aggregation is best illustrated by visual
  8. 8. 480 IEEE TRANSACTIONS ON SYSTEMS. MN AND CYBERNETICS, A. V O L 21. NO. 3, MAYIJUNE 1991images. Temporal aggregation is best illustrated by acousticparameters such as phase, pitch, phonemes, words, sentences,rhythm, beat, and melody. Levels in the world model hierarchy are defined by temporalresolution of events, spatial resolution of maps, and by parent -child relationships between entities in symbolic data structures.These are defined by the needs of both SP and BG modulesat the various levels. Theorem: In a hierarchically structured goal-driven, sensory -interactive, intelligent control system architecture: 1) control bandwidth decreases about an order of magni- tude at each higher level, 2) perceptual resolution of spatial and temporal patterns decreases about an order-of-magnitude at each higher level, 3) goals expand in scope and planning horizons expand in space and time about an order-of-magnitude at each higher level, and 4) models of the world and memories of events decrease in resolution and expand in spatial and temporal range by about an order-of-magnitude at each higher level. I t i s well known from control theory that hierarchicallynested servo loops tend to suffer instability unless the band-width of the control loops differ by about an order o f mag-nitude. This suggests, perhaps even requires, condition 1). Fig. 5. liming diagram illustrating the temporal flow of activity in the taskNumerous theoretical and experimental studies support the decomposition and sensory processing systems. At the world level, high-level sensory events and circadian rhythms react with habits and daily routines toconcept of hierarchical planning and perceptual “chunking” generate a plan for the day. Each elements of that plan is decomposed throughfor both temporal and spatial entities [17], [18]. These support the remaining six levels of task decomposition into action.conditions 2), 3), and 4). In elaboration of the aforementioned theorem, we can con-struct a timing diagram, as shown in Fig. 5. The range of the light shading on the left shows the event summary interval fortime scale increases, and its resolution decreases, exponentiallythe immediately previous about an order of magnitude at each higher level. Hence the Fig. 5 suggests a duality between the behavior generationplanning horizon and event summary interval increases, and and the sensory processing hierarchies. At each hierarchicalthe loop bandwidth and frequency of subgoal events decreases, level, planner modules decompose task commands into stringsexponentially at each higher level. The seven hierarchical of planned subtasks for execution. At each level, strings oflevels in Fig. 5 span a range of time intervals from three sensed events are summarized, integrated, and “chunked” intomilliseconds to one day. Three milliseconds was arbitrarily single events at the next higher level.chosen as the shortest servo update rate because that i s Planning implies an ability to predict future states of theadequate to reproduce the highest bandwidth reflex arc in the world. Prediction algorithms based on Fourier transforms orhuman body. One day was arbitrarily chosen as the longest Kalman filters typically use recent historical data to computehistorical -memory/planning -horizonto be considered. Shorter parameters for extrapolating into the future. Predictions madetime intervals could be handled by adding another layer at the by such methods are typically not reliable for periods longerbottom. Longer time intervals could be treated by adding layers than the historical interval over which the parameters wereat the top, or by increasing the difference in loop bandwidths computed. Thus at each level, planning horizons extend intoand sensory chunking intervals between levels. the future only about as far, and with about the same level of The origin of the time axis in Fig. 5 is the present, i.e., detail, as historical traces reach into the past.t = 0. Future plans lie to the right of t = 0, past history to Predicting the future state of the world often depends onthe left. The open triangles in the right half-plane represent assumptions as to what actions are going to be taken and whattask goals in a future plan. The filled triangles in the left reactions are to be expected from the environment, includinghalf-plane represent recognized task-completion events in a what actions may be taken by other intelligent agents. Planningpast history. At each level there is a planning horizon and a of this type requires search over the space of possible futurehistorical event summary interval. The heavy crosshatching on actions and probable reactions. Search -based planning takesthe right shows the planning horizon for the current task. T h e place via a looping interaction between the BG, WM, and VJlight shading on the right indicates the planning horizon for modules. This is described in more detail in the Section Xthe anticipated next task, The heavy crosshatching on the left discussion on BG modules.shows the event summary interval for the current task. T h e Planning complexity grows exponentially with the number
  9. 9. ALBUS: OUTLINE FOR A THEORY OF INIXLLIGENCE 481 width. Typically the control cycle interval is an order o f magnitude less than the expected output subtask duration. I f the feedback indicates the failure of a planned subtask, the executor branches immediately (i.e., in one control cycle interval) to a preplanned emergency subtask. The planner simultaneously selects or generates an error recovery sequence that is substituted for the former plan that failed. Plan executors are also described in more detail in Section X. When a task goal is achieved at time t = 0, i t becomes a task completion event in the historical trace. To the extent that a historical trace is an exact duplicate of a former plan, there were no surprises; i.e., the plan was followed, and every task was accomplished as planned. To the extent that a historical trace is different from the former plan, there were surprises. The average size and frequency of surprises (i.e., differences between plans and results) is a measure of effectiveness of a t=o planner. At each level in the control hierarchy, the difference vectorFig. 6. Three levels of real-time planning illustrating the shrinking planning between planned (i.e., predicted) commands and observedhorizon and greater detail at successively lower levels of the hierarchy. Atthe top level, a single task is decomposed into a set of four planned subtasks events i s an error signal, that can be used by executorfor each of three subsystem. At each of the next two levels, the first task in submodules for servo feedback control (i.e., error correction),the plan of the first subsystems is further decomposed into four subtasks for and by VJ modules for evaluating success and failure.three subsystems at the next lower level. In the next eight sections, the system architecture out- lined previously wl be elaborated and the functionality of ilof steps in the plan (i.e., the number of layers in the search the computational submodules for behavior generation, worldgraph). If real-time planning i s to succeed, any given planner modeling, sensory processing, and value judgment will bemust operate in a limited search space. I f there are too much discussed.resolution in the time line, or in the space of possible actions,the size of the search graph can easily become too large forreal-time response.. One method of resolving this problem VIII. BEHAVIOR GENERATIONis to use a multiplicity of planners in hierarchical layers[14], [I81 so that at each layer no planner needs to search Definition: Behavior is the result of executing a series ofmore than a given number (for example ten) steps deep in a graph, and at each level there are no more than (ten) Definition: A task is a piece of work to be done, or ansubsystem planners that need to simultaneously generate and activity to be performed.coordinate plans. These criteria give rise to hierarchical levels Axiom: For any intelligent system, there exists a set of taskswith exponentially expanding spatial and temporal planning that the system knows how to do.horizons, and characteristic degrees of detail for each level. Each task in this set can be assigned a name. The taskThe result of hierarchical spatiotemporal planning is illustrated vocabulary is the set of task names assigned to the set of tasksin Fig. 6. At each level, plans consist of at least one, and on the system is capable of performing. For creatures capable ofaverage 10, subtasks. T h e planners have a planning horizon learning, the task vocabulary is not fixed in size. I t can be that extends about one and a half average input command expanded through learning, training, or programming. It may intervals into the future. shrink from forgetting, or program deletion. In a real-time system, plans must be regenerated periodically Typically, a task i s performed by a one or more actors on to cope with changing and unforeseen conditions in the world. one or more objects. T h e performance of a task can usually Cyclic replanning may occur at periodic intervals. Emergency be described as an activity that begins with a start-event and replanning begins immediately upon the detection of an emer- i s directed toward a goal-event. This is illustrated in Fig. 7. gency condition. Under f l alert status, the cyclic replanning u Definition: A goal i s an event that successfully terminates interval should be about an order of magnitude less than a task. A goal is the objective toward which task activity is the planning horizon (or about equal to the expected output directed. subtask time duration). This requires that real-time planners Definirion: A task command is an instruction to perform be able to search to the planning horizon about an order of a named task. A task command may have the form: magnitude faster than real time. This is possible only if the DO <Tasknarne(parameters)> AFTER <Start Event> UNTIL depth and resolution of search is limited through hierarchical <Goal Event> Task knowledge is knowledge of how to planning. perform a task, including information as to what tools, Plan executors at each level have responsibility for react- materials, time, resources, information, and conditions are ing to feedback every control cycle interval. Control cycle required, plus information as to what costs, benefits and risks intervals are inversely proportional to the control loop band- are expected.
  10. 10. 482 IEEE TRANSACTIONS ON SYSTEMS. MAN. AND CYBERNETICS. VOL. 21. NO. 3. MAYIJUNE I V Y 1 Task knowledge is typically difficult to discover, but once known, can be readily transferred to others. Task knowledge may be acquired by trial and error learning, but more often i t is acquired from a teacher, or from written or programmed instructions. For example, the common household task of preparing a food dish is typically performed by following a recipe. A recipe is an informal task frame for cooking. Gourmet dishes rarely result from reasoning about possible combinations o f ingredients, still less from random trial and error combinations o f food stuffs. Exceptionally good recipes / often are closely guarded secrets that, once published, can easily be understood and followed by others. Making steel is a more complex task example. Steel makingFig. 7. A task consists of an activity that typically begins with a start event took the human race many millennia to discover how to do.and is terminated by a goal event. A task may be decomposed into severalconcurrent strings of subtasks that collectively achieve the goal event. However, once known, the recipe for making steel can be implemented by persons of ordinary skill and intelligence. In most cases, the ability to successfully accomplish com- Task knowledge may be expressed implicitly in fixed cir- plex tasks is more dependent on the amount of task knowledgecuitry, either in the neuronal connections and synaptic weights stored in task frames (particularly in the procedure section)of the brain, or in algorithms, software, and computing hard- than on the sophistication of planners in reasoning about tasks.ware. Task knowledge may also be expressed explicitly in datastructures, either in the neuronal substrate or in a computermemory. IX. BEHAVIOR GENERATION Definifion: A task frame is a data structure in which task Behavior generation is inherently a hierarchical process.knowledge can be stored. At each level of the behavior generation hierarchy, tasks are In systems where task knowledge i s explicit, a task frame decomposed into subtasks that become task commands to 9 can be defined or each task i the task f n[1 ] An the next lower level. At each level of a behavior generationexample o f a task frame is: hierarchy there exists a task vocabulary and a corresponding TASKNAME name of the task set of task frames. Each task frame contains a procedure state - type generic or specifi graph. Each node in the procedure state-graph must correspond actor agent performing the task to a task name in the task vocabulary at the next lower level. action activity to be performed thing to be acted upon Behavior generation consists of both spatial and temporal object goal event that successfully terminates or renders the decomposition. Spatial decomposition partitions a task into task successful jobs to be performed by different subsystems. Spatial task parameters priority decomposition results in a tree structure, where each node status (e.g. active, waiting, inactive) timing requirements corresponds to a BG module, and each arc of the tree cor- source of task command responds to a communication link in the chain of command requirements tools, time, resources, and materials needed to as illustrated in Fig. 3. perform the task Temporal decomposition partitions each job into sequential enabling conditions that must be satisfied to begin, subtasks along the time line. T h e result is a set of subtasks, or continue, the task disabling conditions th a t or intempt, all of which when accomplished, achieve the task goal, as the task illustrated in Fig. 7. information that may be required In a plan involving concurrent job activity by different procedures a state-graph or state-table defining a plan for subsystems, there may requirements for coordination, or mu- executing the task tual constraints. For example, a start -event for a subtask functions that may be called algorithms that may be needed activity in one subsystem may depend on the goal-event for eKeets expected results of task execution a subtask activity in another subsystem. Some tasks may expected costs, risks, benefits require concurrent coordinated cooperative action by several estimated time to complete subsystems. Both planning and execution of subsystem plans may thus need to be coordinated. Explicit representation of task knowledge in task frames has There may be several alternative ways to accomplish a task.a variety of uses. For example, task planners may use it for Alternative task or job decompositions can be represented bygenerating hypothesized actions. The may use it an AND/OR graph in the procedure section of the task frame.for predicting the results o f hypothesized actions. T h e value The d e c isio n as t o whi ch o f several alternatives to choose isjudgment system may use i t for computing how important the made through a series of interactions between the BG, WM,goal is and how many resources to expend in pursuing it. Plan SP, and VJ modules. Each alternative may be analyzed by theexecutors may use i t for selecting what to do next. BG module hypothesizing it, WM predicting the result, and VJ
  11. 11. ALBUS: OUTLINE FOR A THEORY OF INELLICENCE 483 as defined by NASREM [14], the JA submodule at each level may also determine the amount and kind of input to accept from a human operator. The Planner Sublevel -PL(j) Submodules j=l,. . : For 2, . N each of the N subsystems, there exists a planner submodule PL(j). Each planner submodule i s responsible for decompos - ing the job assigned to its subsystem into a temporal sequence of planned subtasks. Planner submodules PL(j) may be implemented by case- based planners that simply select partially or completely pre- fabricated plans, scripts, or schema [20]-[22] from the proce - dure sections o f task frames. This may be done by evoking sit - uation/action rules of the form, IF(case -s)/THEN(useglan -y). The planner submodules may complete partial plans by pro- viding situation dependent parameters. The range of behavior that can be generated by a libraryFig. 8. The job assignment JA module performs a spatial decomposition of of prefabricated plans at each hierarchical level, with eachthe task command into N subsystems. For each subsystem, a planner PL(j) plan containing a number o f conditional branches and errorperforms a temporal decomposition o f its assigned job into subtasks. For eachsubsystem, an executor EX(j) closes a real-time control loop that servos the recovery routines, can be extremely large and complex. Forsubtasks to the plan. example, nature has provided biological creatures with an extensive library of genetically prefabricated plans, called instinct. For most species, case-based planning using librariesevaluating the result. The BG module then chooses the “best” of instinctive plans has proven adequate for survival and genealternative as the plan to be executed. propagation in a hostile natural environment. Planner submodules may also be implemented by search- X. BG MODULES based planners that search the space of possibile actions. This In the control architecture defined in Fig. 3, each level of requires the evaluation o f alternative hypothetical sequencesthe hierarchy contains one or more BG modules. At each level, of subtasks, as illustrated in Fig. 9. Each planner PL(j)there is a BG module for each subsystem being controlled. The hypothesizes some action or series of actions, the W M modulefunction of the BG modules are to decompose task commands predicts the effects of those action(s), and the VJ moduleinto subtask commands. computes the value of the resulting expected states of the Input to BG modules consists of commands and priorities world, as depicted in Fig. 9(a). This results in a game (orfrom BG modules at the next higher level, plus evaluations search) graph, as shown in 9(b). The path through the gamefrom nearby VJ modules, plus information about past, present, graph leading to the state with the best value becomes the planand predicted future states of the world from nearby W M to be executed by EX(j). In either case-based or search -basedmodules. Output from BG modules may consist of subtask planning, the resulting plan may be represented by a state-commands to BG modules at the next lower level, plus status graph, as shown in Fig. s(~). Plans may also be representedreports, plus “What Is?” and “What I ? ueries to the WM f ” q by gradients, or other types of fields, on maps [23], or inabout the current and future states of the world. configuration space. Each BG module at each level consists o f three sublevels Job commands to each planner submodule may contain[9], [14] as shown in Fig. 8. constraints on time, or specify job-start and job-goal events. The Job Assignment Sublevel -JA Submodule: The JA sub- A job assigned to one subsystem may also require synchro -module i s responsible for spatial task decomposition. I t par- nization or coordination with other jobs assigned to different titions the input task command into N spatially distinct jobs subsystems. These constraints and coordination requirements to be performed by N physically distinct subsystems, where may be specified by, or derived from, the task frame. Each N i s the number of subsystems currently assigned to the BG planner PL(j) submodule i s responsible for coordinating module. The JA submodule many assign tools and allocate its plan with plans generated by each of the other N - 1 physical resources (such as arms, hands, legs, sensors, tools, planners at the same level, and checking to determine if and materials) to each of its subordinate subsystems for their there are mutually conflicting constraints. I f conflicts are use in performing their assigned jobs. These assignments are found, constraint relaxation algorithms [24] may be applied, not necessarily static. For example, the job assignment sub- or negotiations conducted between PL(j) planners, until a module at the individual level may, at one moment, assign an solution i s discovered. If no solution can be found, the planners arm to the manipulation subsystem i n response to a cusetoob report failure to the job assignment submodule, and a new job task command, and later, assign the same arm to the attention assignment may be tried, or failure may be reported to the subsystem in response to a ctouch/feel> task command. next higher level BG module. T h e job assignment submodule selects the coordinate sys- The Executor Sublevel-EX(j) Submodules: There i s an ex- tem in which the task decomposition at that level is to be ecutor EX(j) for each planner PL(j). The executor sub- performed. In supervisory or telerobotic control systems such modules are responsible for successfully executing the plan
  12. 12. 484 IEEE TRANSACnONS ON SYSTEMS, MAN. AND CYBERNFTICS. VOL. 21. NO. 3. MAYllUNE 1991 there exists a hierarchy of task vocabularies that can be overlaid on the spatial/temporal hierarchy of Fig. 5. For example: Level 1 is where commands for coordinated velocities and forces of body components (such as arms, hands, fingers, legs, eyes, torso, and head) are decomposed into motor commands to individual actuators. Feedback servos the position, velocity, and force of individual actuators. In vertebrates, this is the level of the motor neuron and stretch reflex. Level 2 is where commands for maneuvers o f body com- ponents are decomposed into smooth coordinated dynamically efficient trajectories. Feedback servos coordinated trajectory motions. This is the level o f the spinal motor centers and the cerebellum. Level 3 is where commands to manipulation, locomotion, and attention subsystems are decomposed into collision free paths that avoid obstacles and singularilies. Feedback servos movements relative to surfaces in the world. This is the level of the red nucleus, the substantia nigra, and the primary motor cortex.Fig. 9. Planning loop (a) produces a game graph (b). A trace in the gamegraph from the Stan to a goal state is a plan that can be represented as a plan Level 4 is where commands for an individual to performgraph (c). Nodes in the game graph correspond to edges in the plan graph, simple tasks on single objects are decomposed into coordi -and edges in the game graph correspond to nodes in the plan graph. Multiple nated activity o f body locomotion, manipulation, attention, andedges exiting nodes in the plan graph correspond to conditional branches. communication subsystems. Feedback initiates and sequences subsystem activity. This is the level of the basal ganglia andstate-graphs generated by their respective planners. At each pre-motor frontal cortex.tick o f the state clock, each executor measures the difference Level 5 is where commands for behavior of an intelligentbetween the current world state and its current plan subgoal self individual relative to others in a small group are decom -state, and issues a subcommand designed to null the difference. posed into interactions between the self and nearby objects orWhen the world model indicates that a subtask in the current agents. Feedback initiates and steers whole self task activity.plan is successfully completed, the executor steps to the next Behavior generating levels 5 and above are hypothesized tosubtask in that plan. When all the subtasks in the current reside in temporal, frontal, and limbic cortical areas.plan are successfully executed, the executor steps to the first Level 6 is where commands for behavior of the individualsubtask in the next plan. If the feedback indicates the failure relative to multiple groups are decomposed into small groupof a planned subtask, the executor branches immediately to a interactions. Feedback steers small group interactions.preplanned emergency subtask. Its planner meanwhile begins Level 7 (arbitrarily the highest level) is where long rangework selecting or generating a new plan that can be substi- goals are selected and plans are made for long range behaviortuted for the former plan that failed. Output subcommands relative to the world as a whole. Feedback steers progressproduced by executors at level i become input commands to toward long range goals.job assignment submodules in BG modules at level i - 1. The mapping o f BG functionality onto levels one to four Planners PL(j) operate on the future. For each subsystem, defines the control functions necessary to control a singlethere i s a planner that i s responsible for providing a plan intelligent individual in performing simple task goals. Func-that extends to the end of its planning horizon. Executors tionality at levels one through three is more or less fixed andEX(j) operate in the present. For each subsystem, there i s an specific to each species of intelligent system [25]. At levelexecutor that is responsible for monitoring the current (t = 0) 4 and above, the mapping becomes more task and situationstate of the world and executing the plan for its respective dependent. Levels 5 and above define the control functionssubsystem. Each executor performs a READ -COMPUTE - necessary to control the relationships of an individual relativeWRITE operation once each control cycle. At each level, each to others in groups, multiple groups, and the world as a whole.executor submodule closes a reflex arc, or servo loop. Thus, There i s good evidence that hierarchical layers develop inexecutor submodules at the various hierarchical levels form a the sensory -motor system, both in the individual brain as theset o f nested servo loops. Executor loop bandwidths decrease individual matures, and in the brains of an entire species as theon average about an order of magnitude at each higher level. species evolves. I t can be hypothesized that the maturation of levels in humans gives rise to Piaget’s “stages of development” (261. Of course, the biological motor system is typically much XI. THE B EH ~VIOR ENERATING HIERARCHY G more complex than is suggested by the example model de- Task goals and task decomposition functions often have scribed previously. In the brains of higher species there maycharacteristic spatial and temporal properties. For any task, exist multiple hierarchies that overlap and interact with each
  13. 13. ALBUS: OUTLINE FOR A THEORY OF INTELLIGENCE 485 other in complicated ways. For example in primates, the Value Judgmcnl pyramidal cells of the primary motor cortex have outputs to the motor neurons for direct control of fine manipulation as well as the inferior olive for teaching behavioral skills Task to the cerebellum [27]. There is also evidence for three Planner parallel behavior generating hierarchies that have developed over three evolutionary eras [28]. Each BG module may thus contain three or more competing influences: 1) the most basic Sensory Task Compare Executor (IF i t smells good, THEN eat it), 2) a more sophisticated(WAIT until the “best” moment) where best is when successprobability i s highest, and 3) a very sophisticated (WHAT are the long range consequences of my contemplated action, andwhat are all my options). On the other hand, some motor systems may be less complexthan suggested previously. Not all species have the samenumber of levels. Insects, for example, may have only two orthree levels, while adult humans may have more than seven. In Urobots, the functionality required of each BG module dependsupon the complexity of the subsystem being controlled. For Fig. 10. Functions performed by the WM module. 1) Update knowledge database with prediction errors and recognized entities. 2) Predict sensoryexample, one robot gripper may consist o f a dexterous hand data. 3) Answer “What is?” queries from task executor and return currentwith 15 to 20 force servoed degrees of freedom. Another state of world. 4) Answer “What if?” queries from task planner and predictgripper may consist of two parallel jaws actuated by a single results for evaluation.pneumatic cylinder. In simple systems, some BG modules(such as the Primitive level) may have no function (such A. WM Modulesas dynamic trajectory computation) to perform. In this case, The WM modules in each node of the organizational hi-the BG module will simply pass through unchanged input erarchy of Figs. 2 and 3 perform the functions illustrated incommands (such as <Grasp>). Fig. 10. W M modules maintain the knowledge database, keeping XII. THE WORLD MODEL i t current and consistent. In this role, the WM modules Definition: T h e world model is an intelligent system’s perform the functions of a database management system.internal representation of the external world. I t is the system’s They update WM state estimates based on correlationsbest estimate o f objective reality. A clear distinction between and differences between world model predictions andan internal representation of the world that exists in the sensory observations at each hierarchical level. Themind, and the external world of reality, was first made in WM modules enter newly recognized entities, states,the West by Schopenhauer over 100 years ago [29]. In the and events into the knowledge database, and deleteEast, i t has been a central theme of Buddhism for millennia. entities and states determined by the sensory processingToday the concept of an internal world model is crucial modules to no longer exist in the external world. T h eto an understanding of perception and cognition. The world WM modules also enter estimates, generated by the VJmodel provides the intelligent system with the information modules, of the reliability of world model state variables.necessary to reason about objects, space, and time. The world Believability or confidence factors are assigned to manymodel contains knowledge of things that are not directly and types of state variables.immediately observable. I t enables the system to integrate WM modules generate predictions of expected sensorynoisy and intermittent sensory input from many different input for use by the appropriate sensory processingsources into a single reliable representation of spatiotemporal SP modules. In this role, a W M module performs thereality. functions of a signal generator, a graphics engine, or Knowledge in an intelligent system may be represented state predictor, generating predictions that enable theeither implicitly or explicitly. Implicit world knowledge may sensory processing system to perform correlation andbe embedded in the control and sensory processing algorithms predictive filtering. W M predictions are based on theand interconnections of a brain, or of a computer system. state o f the task and estimated states of the externalExplicit world knowledge may be represented in either natural world. For example in vision, a W M module may useor artificial systems by data in database structures such as the information in an object frame to generate real -time maps, lists, and semantic nets. Explicit world models require predicted images that can be compared pixel by pixel, computational modules capable of map transformations, indi- or entity by entity, with observed images. rect addressing, and list processing. Computer hardware and 3) W M modules answer “What is?” questions asked by the software techniques for implementing these types of functions planners and executors in the corresponding level BG are well known. Neural mechanisms with such capabilities are modules. In this role, the W M modules perform the func- discussed in Section XVI. tion of database query processors, question answering